Yes — with 45.6 GB to spare
Qwen3 1.7B at Q4_K_M fits your M3 Max · 64 GB entirely in unified memory at 8K context, at an estimated 231 tokens per second. There is room for its full 32K window.
Fully in unified memory
8K context
Q4_K_M · 1.0 GB
Apache 2.0
Released Apr 2025
Punches above its size on structured tasks, with optional thinking mode.
The VRAM budget
weights 1.0 GB
Weights 1.0 GB
KV cache @ 8K 0.88 GB
Runtime overhead 0.6 GB
Free 45.6 GB of 48.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 3.2 GB | 4.7 GB | 32K | 70 | Reference | Long context |
| Q8_0 | 1.7 GB | 3.2 GB | 32K | 131 | −0.1% ppl | Long context |
| Q6_K | 1.3 GB | 2.8 GB | 32K | 170 | −0.4% ppl | Long context |
| Q5_K_M | 1.1 GB | 2.6 GB | 32K | 197 | −0.8% ppl | Long context |
| Q4_K_M | 1.0 GB | 2.4 GB | 32K | 231 | −1.9% ppl | Recommended |
| Q3_K_M | 0.8 GB | 2.3 GB | 32K | 285 | −5.4% ppl | Long context |
| Q2_K | 0.7 GB | 2.1 GB | 32K | 333 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Qwen3-1.7B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 1.0 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 48.0 GB of 64 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 32K context on this card.